Computer Science · Computational Theory and Mathematics
Subspace Iteration for Eigenproblems
Henk Van Der, Henk Van Der Vorst
See on the mapAbstract
Introduction The computation of eigenvalues and eigenvectors for the standard eigenproblem was considered a solved problem 15 years ago, as far as matrices of modest order were concerned. The situation for large matrices, say of order n = 2; 000 or more, was less satisfactory, since these matrices can not be treated by the standard direct approaches: reduction to special form, plus QR-iteration on this special form. This requires O(n 3 ) arithmetic operations, and apart from storage considerations, it is clear that this puts limits on n. For very large sparse symmetric matrices, the iteration method of Lanczos [9] was more or less accepted around 1980, due to pioneering work of Paige [10]. Various effects in the convergence of eigenvalue approximations in the Lanczos process have been analysed in great detail in [16]. Similar methods for unsymmetric matrices were available, but their beha